30th Amateur Series Division 1

[ernest]

Komodo 3.0 32-bit is almost exactly 2 times slower than the 64-bit (no SSE...), which is quite a record among engines

[Graham Banks]

Komodo 3.0 32-bit is almost exactly 2 times slower than the 64-bit (no SSE...), which is quite a record among engines

[Graham Banks]

30TH AMATEUR SERIES (Division 1)

Intel i5 750
ChessGUI
128mb hash each
3-4-5 piece tablebases
Ponder off
Charlotte 1.0 mini.cgb book
40 moves in 25 minutes repeating (adapted for the CCRL)
4 cycles 44 rounds
All engines 32-bit 1CPU

The bottom engine will face automatic relegation. Whether or not the second from bottom engine (and perhaps others) remain in this division could be determined by grading matches before the next series.

Final Standings

35.5 - Houdini 1.5a 32-bit
33.0 - Critter 1.2 32-bit
28.5 - Stockfish 2.1.1 32-bit
27.5 - Komodo 3.0 32-bit
22.5 - Spike 1.4 Leiden
19.5 - Booot 5.1.0
19.0 - Thinker 5.4d Inert 32-bit
17.5 - Protector 1.4.0 32-bit
17.5 - Hannibal 1.1 32-bit
17.0 - Spark 1.0 32-bit
15.0 - Umko 1.2 32-bit
11.5 - Jonny 4.00


The complete tournament pgn (zipped) can be downloaded here:
http://kirill-kryukov.com/chess/discuss ... p?id=23506

[Frank Quisinsky]

Hi Ernest,

after the SWCR-32 Update tourney (end in around 2-3 weeks) I have for over 50 engine versions the 32bit to 64bit compare. I don't know how you calculate the factor you wrote about on more times in TalkChess. I think it's an easy and locical calculation with perhaps 3-5 positions (middlegame 2 and endgame 3).

It's possible for yourself to calculate for this 50 engine versions again the differents. So we can create a nice HTML about it, perhaps with a short text how you do that?

If you have interest on it ...

Best
Frank

[ernest]

I have for over 50 engine versions the 32bit to 64bit compare.

Hi Frank,

For every new engine or version, which comes in 32-bit and 64-bit,
I compare the infinite analysis of the start_position, for the 32-bit and 64-bit versions/compiles.
I use 512MB hash and 1 thread for reproducibility.

Mostly (but there are exceptions), the 32-bit and 64-bit versions give the same analysis (same kNodes, same eval and PV for a given depth), only the times are different. So it is easy to compute the speed ratio.

When the analysis is not the same, you get reproducible kNodes/sec, which also gives you a speed ratio.

Of course, if the infinite analysis is not of the start_position, but of some middle_game or end_game position, the ratio will be different, but should not be too different (except maybe completely at the end, when tablebases are in action)

I have checked that the 64-bit/32-bit ratio is a good indicator of the Elo difference between the 64-bit and 32-bit versions. For instance, Rybka has a x1.7 ratio, corresponding to about +50 Elo

[Frank Quisinsky]

Hi Ernest,

thanks for your clear answer!
OK, that explain all, idea is good!

It would be nice to have an excel with the different in ratings from 32bit to 64bit SWCR results in column 1 and the factor you wrote about it in column 2 with a column 3 for the differences. I think you are right in the compare you do but it could be possible that for different engines the differences are bigger / smaler in your or my test, I don't know.

A nice idea I believe to create such an excel. For my ratinglist is the difference from 32bit to 64bit a main point. Many readers of my site using 32bit OS or different machines, with 32bit or 64bit OS.

Thanks again!
Have a nice day!

Best
Frank